Fold hourly 1454 HIGH into notes §153

[24601]

Summary

Hourly 1454 fold off current main (c8785c4, ryana/jevify §152 / merged #78). This fold is notes.md §153, composition 821-836, findings batch #133. It does not reclaim §147 through §152 and does not reopen PR #23 through #78.

TypeSafe Jev stays the default path. The class mandate stays: classifiers, encoders and decoders, specialized AR and constrained heads, vision and listwise scorers, and what TypeSafe calls System One.

• NobleSpartan6/otto. Jev chooses concrete actions. Code acts after review. Fill-a-form is exact. A model score is not a guarantee of success.
• nourhelmi/pi-jev-compaction. Clear only below 0.25. API failure clears nothing new.
• LiamSherline/jev-lead-scorer. Nothing sends itself. P near 0.5 is NULL.
• MatteoGauthier/laya-portable. "Same answers" is their claim. Export parity is not Harbor.
• strombolini/Armada. Codiv OpenJev ranking is not calibration and is not TypeSafe.
• Class benches stay theirs (84.0%, 88.58%, one Mind2Web shard, n=40).
• hf:lostargon/Tiny-Jev. In-distribution ECE 0.004 theirs is not the held-out ECE 0.299 theirs. Fitted public rows are not a zero-shot comparison.
• REVISIT densifies stay on the prior sections: swev §132 (MLX was CoreML), zero-shot-classifier §145, and §151 for rarha/laya-onnx, TokyoHunter, and aleksvega.

The uniqueness lock lives in research/notes.md and the gate fixture only. README still ends at License. Does not bump 0.5.0 or 0.5.1. invented_signal: false.

Test plan

• python3 .agents/skills/augustus/scripts/uniqueness_gate.py
• python3 research/revisit_fingerprints.py --self-test
• Adversarial pass: no em dashes in the new prose, theirs on third-party numbers, namesakes kept apart, soft judgment is not a sole veto. Tiny-Jev card corrected after the first pass so the held-out ECE is not dropped.